Nested Cross-Validation Reveals Performance Inflation in MRI Radiomics for Early Mortality Prediction in IDH-Wildtype...

Background: Glioblastoma (GBM) carries a median survival of 15 months. Approximately 16-27% of patients die within six months despite standard therapy. Radiomic MRI features have been proposed as prognostic biomarkers, yet many published studies employ standard cross-validation (CV) with feature selection on the full dataset, introducing data leakage. Methods: Sixty patients with IDH-wildtype GBM from the publicly available UCSF-PDGM dataset (single-center, GE Discovery MR750, 3 T; 2015-2021) with BraTS 2021 segmentation masks were analyzed. One thousand two hundred eighty-four (1284) IBSI-compliant radiomic features (7 feature classes × 4 MRI sequences × 3 tumor subregions) were extracted using PyRadiomics. After preprocessing (variance filter, correlation filter |r| > 0.95), approximately 120 features remained per fold. Feature selection was performed strictly within each training fold of a nested 5-fold cross-validation framework (5-fold × 3 repeats = 15 outer folds). Results: For 1-year mortality, radiomics AUC dropped from 0.816 (standard CV) to 0.593 (nested CV; ΔAUC = -0.223, 27% inflation), while clinical models remained stable (0.708 vs. 0.705). For early mortality (≤180 days, n = 16 events), standard CV inflated radiomics AUC to 0.888, whereas nested CV yielded 0.815 (ΔAUC = -0.073, 8% inflation). In feature stability analysis, whole-tumor surface area (13/15 folds) and mesh volume (10/15 folds) showed the highest cross-endpoint stability; nine of 13 exploratory OS-associated features were never selected for early mortality classification. Extent of resection (HR = 0.44, p = 0.009) and age (HR = 1.03, p = 0.038) were independently associated with overall survival; tumor surface area remained independently associated with survival (HR = 1.38, p = 0.021). Conclusions: Nested cross-validation revealed substantial performance inflation in standard radiomics pipelines. These descriptive inflation estimates are specific to this dataset and pipeline configuration and should not be generalized as universal parameters for radiomics.